基于MA-DRNet的糖尿病视网膜病变等级识别方法

糖尿病视网膜病变是糖尿病并发症最常见的疾病之一。由于视网膜病变病灶具有特征复杂、特征差异小的特点，导致传统深度学习网络对视网膜病变等级识别存在错误率高、鲁棒性差等问题。针对上述问题，提出了一种MA-DRNet模型进行优化：(1)提出了一种多级特征残差块，提取不同分辨率多尺度特征、扩大模型感受野，加强模型对于小尺度病灶的学习能力以及模型对尺度的鲁棒性；(2)改进一种全局通道联合注意力机制，实现像素长距离依赖关系捕获和通道注意力，提升模型对复杂病灶表征效果；(3)设计集成难例挖掘训练方法，巩固对于困难样本的学习，融入集成的思想提升模型对易错样本的关注度。在Kaggle和Messidor两个公开视网膜数据集进行模型训练和测试，本文模型特异性为99.02%,敏感性为98.26%,准确率为98.87%,各指标均优于目前同类算法。大量的实验表明，MA-DRNet有效的解决了视网膜病变识别存在的问题，实现了视网膜病变等级的高精度辅助诊断。

Research on Level Recognition Method of Diabetic Retinopathy Based on MA-DRNet

Diabetic retinopathy is one of the most common complications of diabetes. Due to the characteristics of complex features and small differences in features of retinopathy lesions, the traditional deep learning network has problems such as high error rate and poor robustness in the recognition of retinopathy grades. In view of the above problems, this paper proposes a MA-DRNet model for optimization: (1) A multi-level feature residual block is proposed to extract multi-scale features of different resolutions, expand the receptive field of the model, and strengthen the model''s sensitivity to small-scale lesions. Learning ability and the robustness of the model to scale; (2) Improve a global channel joint attention mechanism to achieve pixel long-distance dependency capture and channel attention, and improve the model''s representation of complex lesions; (3) The training method for ensemble hard example mining is designed to consolidate the learning of difficult samples, and integrate the idea of to improve the model''s attention to error-prone samples. The model is trained and tested on two public retinal datasets, Kaggle and Messidor. The model in this paper has a specificity of 98.96%, a sensitivity of 97.86%, and an accuracy of 98.82%. All indicators are better than the current similar algorithms. A large number of experiments show that MA-DRNet effectively solves the problem of retinopathy identification and achieves high-precision auxiliary diagnosis of retinopathy grade.